File size: 7,236 Bytes
7af453b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 | # coding=utf-8
"""Raw 8-bit byte tokenizer for x8D-Omni-Diffusion.
Pure Python standard library only. No torch, no transformers, no BPE,
no vocab.json, no merges.txt.
Vocabulary is the 256 unsigned byte states (0x00-0xFF) plus 8 special
tokens:
MASK=256, PAD=257, BOS=258, EOS=259,
IMG_START=260, IMG_END=261, AUD_START=262, AUD_END=263
Total vocab size = 264.
"""
from __future__ import annotations
from typing import Dict, Iterable, List, Union
VOCAB_SIZE = 264
MASK_TOKEN_ID = 256
PAD_TOKEN_ID = 257
BOS_TOKEN_ID = 258
EOS_TOKEN_ID = 259
IMG_START_TOKEN_ID = 260
IMG_END_TOKEN_ID = 261
AUD_START_TOKEN_ID = 262
AUD_END_TOKEN_ID = 263
SPECIAL_IDS = frozenset(
{
MASK_TOKEN_ID,
PAD_TOKEN_ID,
BOS_TOKEN_ID,
EOS_TOKEN_ID,
IMG_START_TOKEN_ID,
IMG_END_TOKEN_ID,
AUD_START_TOKEN_ID,
AUD_END_TOKEN_ID,
}
)
#: Map a human-friendly name to the special token id.
SPECIAL_TOKEN_ID_MAP: Dict[str, int] = {
"mask": MASK_TOKEN_ID,
"pad": PAD_TOKEN_ID,
"bos": BOS_TOKEN_ID,
"eos": EOS_TOKEN_ID,
"img_start": IMG_START_TOKEN_ID,
"img_end": IMG_END_TOKEN_ID,
"aud_start": AUD_START_TOKEN_ID,
"aud_end": AUD_END_TOKEN_ID,
}
def _to_bytes(data: Union[str, bytes, bytearray, Iterable[int]]) -> bytes:
"""Normalize any accepted input type into a bytes object."""
if isinstance(data, bytes):
return data
if isinstance(data, bytearray):
return bytes(data)
if isinstance(data, str):
return data.encode("utf-8")
return bytes(int(b) & 0xFF for b in data)
class ByteTokenizer:
"""Encode/decode arbitrary byte streams with a 264-id vocabulary.
The tokenizer performs NO vocabulary lookup and NO sub-word merging.
Text, images, audio, code, and binaries all reduce to the same raw
byte array: ``list(data_bytes)``.
"""
vocab_size: int = VOCAB_SIZE
mask_token_id = MASK_TOKEN_ID
pad_token_id = PAD_TOKEN_ID
bos_token_id = BOS_TOKEN_ID
eos_token_id = EOS_TOKEN_ID
img_start_token_id = IMG_START_TOKEN_ID
img_end_token_id = IMG_END_TOKEN_ID
aud_start_token_id = AUD_START_TOKEN_ID
aud_end_token_id = AUD_END_TOKEN_ID
def encode(
self,
data: Union[str, bytes, bytearray, Iterable[int]],
add_special_tokens: bool = True,
) -> List[int]:
"""Convert input to a list of byte ids in [0, 255].
With ``add_special_tokens=True`` the stream is wrapped as
``[BOS] ... [EOS]``.
Args:
data: text (UTF-8 encoded), raw bytes, or an iterable of ints.
add_special_tokens: whether to prepend BOS and append EOS.
Returns:
List of integer ids. Content ids are always 0-255.
"""
ids = list(_to_bytes(data))
if add_special_tokens:
return [BOS_TOKEN_ID, *ids, EOS_TOKEN_ID]
return ids
def encode_image(
self,
image_bytes: Union[bytes, bytearray, Iterable[int]],
add_special_tokens: bool = True,
) -> List[int]:
"""Encode raw image pixel bytes wrapped in IMG_START/IMG_END."""
ids = list(_to_bytes(image_bytes))
wrapped = [IMG_START_TOKEN_ID, *ids, IMG_END_TOKEN_ID]
if add_special_tokens:
return [BOS_TOKEN_ID, *wrapped, EOS_TOKEN_ID]
return wrapped
def encode_audio(
self,
pcm_bytes: Union[bytes, bytearray, Iterable[int]],
add_special_tokens: bool = True,
) -> List[int]:
"""Encode raw PCM/audio bytes wrapped in AUD_START/AUD_END."""
ids = list(_to_bytes(pcm_bytes))
wrapped = [AUD_START_TOKEN_ID, *ids, AUD_END_TOKEN_ID]
if add_special_tokens:
return [BOS_TOKEN_ID, *wrapped, EOS_TOKEN_ID]
return wrapped
def decode(
self,
ids: Iterable[int],
skip_special_tokens: bool = True,
as_bytes: bool = True,
) -> Union[bytes, str]:
"""Convert ids back into raw bytes or UTF-8 text.
Content ids (0-255) are mapped directly back to bytes. Special ids
are dropped when ``skip_special_tokens`` is True (default).
Args:
ids: iterable of integer ids.
skip_special_tokens: drop ids >= 256 when True.
as_bytes: return ``bytes`` when True, else decode as UTF-8 str.
Returns:
The reconstructed byte string, or its UTF-8 text.
"""
raw = bytearray()
for i in ids:
i = int(i) & 0x1FF
if skip_special_tokens and i in SPECIAL_IDS:
continue
if i < 256:
raw.append(i)
if as_bytes:
return bytes(raw)
return bytes(raw).decode("utf-8", errors="replace")
def encode_text(self, text: str, add_special_tokens: bool = True) -> List[int]:
"""Convenience: encode a UTF-8 string."""
return self.encode(text, add_special_tokens=add_special_tokens)
def decode_text(self, ids: Iterable[int], skip_special_tokens: bool = True) -> str:
"""Convenience: decode ids into a UTF-8 string."""
return self.decode(ids, skip_special_tokens=skip_special_tokens, as_bytes=False)
def __call__(
self,
text: Union[str, bytes, bytearray, Iterable[int]],
add_special_tokens: bool = True,
) -> Dict[str, List[int]]:
"""Tokenize as a dict for drop-in pipeline compatibility.
Returns ``{"input_ids": [...], "attention_mask": [...]}``.
"""
ids = self.encode(text, add_special_tokens=add_special_tokens)
return {
"input_ids": ids,
"attention_mask": [1] * len(ids),
}
def __len__(self) -> int:
return VOCAB_SIZE
def get_vocab(self) -> Dict[int, str]:
"""Return a complete 264-entry id -> label mapping.
Bytes 0-255 map to a printable hex label; specials map to their
token name. Exists for tooling/debugging only -- never used for
encoding (there is no vocabulary lookup in this framework).
"""
vocab: Dict[int, str] = {}
for b in range(256):
vocab[b] = f"byte({b:02X})"
for name, i in SPECIAL_TOKEN_ID_MAP.items():
vocab[i] = f"<{name}>"
return vocab
def bytes_to_ids(self, data_bytes: bytes) -> List[int]:
"""The fundamental law: ``list(data_bytes)`` with no encoding step."""
return list(data_bytes)
def id_batch(self, sequences: Iterable[Iterable[int]]) -> List[List[int]]:
"""Left-pad a batch to equal length with PAD (id 257)."""
seqs = [list(s) for s in sequences]
max_len = max(len(s) for s in seqs) if seqs else 0
return [
[PAD_TOKEN_ID] * (max_len - len(s)) + list(s)
for s in seqs
]
@staticmethod
def mask_canvas(seq_len: int) -> List[int]:
"""Produce a fully-masked diffusion canvas of MASK (id 256)."""
return [MASK_TOKEN_ID] * seq_len
@staticmethod
def renoise_to_random_bytes(positions: Iterable[int]) -> List[int]:
"""Uniform-state re-noise: rejected positions become random bytes."""
return [b & 0xFF for b in positions]
|